Raki's notes on reading paper: named entity recognition as dependency parsing

Abstract & Introduction & Related Work

• Research tasks
  nesting NER
• Existing methods and related work
• Facing the challenge
• Innovative ideas
  Use dependency resolution to do NER
• The experimental conclusion
  sota

Methods

For input bert,fasttext,char Three embedding

bert use (Kantor and Globerson, 2019) The plan , Get a goal token Context sensitive embedding of , On each side 64 All around token

char embedding use CNN

Three embedding concat Get up and throw it BiLSTM Inside

For each of these span, Put the first and last features into a full connection layer , obtain h And then we use double affine to get $r_m$

$r_m$ It provides all possible under the condition that the head is in front of the tail span A score that can constitute a named entity , And then through argmax Get an entity category

For nested NER, As long as an entity does not conflict with the boundaries of higher ranked entities , Will be selected

For a plane NER, We also apply a constraint , That is, any entity that contains or precedes it will not be selected . The learning goal of our named entity recognizer is to assign to a correct category （ Including non entities ） Give each valid span. therefore , It is a multi class classification problem , We use it softmax Cross entropy to optimize our model

Experiments

Conclusion

In this paper , We will NER Reformulated as a structured forecasting task , And adopted SoTA The dependency resolution method is used for nesting and flat NER. Our system uses context embedding as a multi-layer BiLSTM The input of . We use a bilinear model to assign scores to all spans in a sentence . Further constraints are used to predict nested or planar named entities . We evaluated our system in eight named entity corpora . Results show , Our system has been implemented in all eight corpora SoTA. We demonstrate the effectiveness of advanced structured prediction techniques for nested and planar NER Have greatly improved

Remark

I had been dissuaded from seeing the title , Dependency resolution ？？ I can't ,CS224n I was so confused about this that I didn't make it clear , Then look at the model , Only one page ？？！ Start reading now , Then finish reading , I feel refreshed , It doesn't seem to have much to do with dependency resolution , With a simple model sota, Not the kui is a Google Participate in production , It's much better than the last one , But also the research is popular flat/nested NER, I love you